model: which image model to use; omit for the configured default. Currently available: "qwen21"
    (Qwen Image 2.1). list_models shows what is installed and ready.
    Canvas: give width and height (multiples of 32, default 1024×1024) or an aspect preset:
    square, landscape (1344×768), portrait (768×1344), wide (1536×640), tall (640×1536).
    seed: omit for random; the seed used is always returned so a result can be reproduced.
    steps default 25, cfg default 1.0 (Qwen 2.1 works best at cfg 1.0; a negative prompt only
    matters when cfg > 1). count 1-4 images in one run.
    save=true writes full-size PNGs to the configured output folder and returns their paths plus
    a small preview of each; save=false returns base64 PNG in the text instead of writing files.
    Blocks until done (typically 20-60 s). If it returns JSON with a job_id, the server was busy:
    use wait_for_job(job_id).
    

    model: which image model to use; omit for the configured default. Currently available: "qwen21".
    images: 1-4 local file paths (or data:/base64 strings). ORDER MATTERS: the first image is
    the base/subject and sets the output canvas (unless width/height are given); it is fitted and
    padded, never stretched. Later images are supporting references at ~1 MP.
    instruction: start with the operation ("Replace the background with…", "Put the person from
    <image1> into the scene from <image2>…"). Say what comes from each reference, what must stay
    recognizable and what must change. For placing a subject into a scene: subject FIRST, scene
    SECOND, with explicit <image1>/<image2> roles.
    Other parameters and the return value match generate_image.